Agentic AI for the Next Generation of Financial Advisors

Today, client executives working with corporate clients use a variety of software tools and data sources to provide insights to their clients. Time-consuming analysis, limited availability of internal experts, and difficulty accessing the right data limits the ability to provide meaningful insights to the client.

Inspired by SEB joining the consortium building the largest AI Enterprise supercomputer in Sweden, in late 2025, we started working together with NVIDIA Technology Center on building a system for client executives. Our ambition is to enable client executives, primarily working with large SMEs and Midcorp companies, to provide qualitatively better advice – leading to retaining and attracting new clients. We presented the first outcomes of this project at NVIDIA GTC on March 18th.

Agentic technology enabling deeper analysis

Architecture at glance

We chose to build an agentic system bringing together various data sources, including SEB's own research reports provided by our internal experts, as well as internal and external information about companies.

Inspired by NVIDIA AI-Q blueprint blueprint, we built the first module of our system, analysing how the macroeconomic situation can affect a specific company - Global Market Outlook. In the next phases of the project, we plan to extend the functionality with a number of specialized agents analysing various aspects of the company's present and expected future performance, running scenario analysis, and incorporating predictive modelling to enable proactive advice. We chose a flexible open-source framework - Agent Development Kit - to build our system, as it allows us to easily switch between proprietary and open-source GenAI components, enabling interoperability between Google Cloud Platform and AI Factory.

Global Market Outlook Architecture

Sovereign AI: using NVIDIA NIM microservices to replace proprietary models

One of the crucial components of the system is the RAG pipeline. NVIDIA NeMo Retriever provides a competitive alternative to proprietary RAG solutions, allowing us to deploy it on AI Factory and have full control of our data. Furthermore, unlike black-box solutions, NeMo Retriever can be further customized to improve the performance. As data extraction from documents and RAG is one of the most common architecture patterns in a large number of use cases across the bank, we envision using the pipeline deployed on AI Factory to enable dozens of GenAI applications.

Once AI Factory is operational, we are planning to deploy open-source GenAI components, and specifically the NeMo Retriever pipeline, there, and conduct scalability studies to ensure low latency independently of the number of concurrent users.

Anastasia Varava

Head of Research, SEBx